AN ANALYSIS OF TWO COMMON REFERENCE POINTS FOR EEGS.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Data Science

Background:

• Clinical electroencephalographic (EEG) data exhibits significant variability due to operational conditions like electrode type and placement.
• Two common referential montages, Linked Ear (LE) and Averaged Reference (AR), each constitute a substantial portion (approx. 45%) of the TUH EEG Corpus.
• Understanding and mitigating this variability is crucial for reliable EEG data analysis.

Purpose of the Study:

• To investigate the statistical differences between LE and AR EEG data montages.
• To evaluate the impact of these statistical differences on machine learning model performance.
• To identify optimal preprocessing strategies for improving classification accuracy.

Main Methods:

• Comparison of statistical properties of features extracted from LE and AR EEG data.
• Training and evaluation of a Hidden Markov Model (HMM) based classification system on LE, AR, and combined datasets.
• Analysis of preprocessing techniques including mean, variance, channel normalization, and cepstral mean subtraction.

Main Results:

• A HMM system trained solely on LE data achieved significantly higher performance (77.2%) compared to one trained only on AR data (61.4%).
• Training a system on both LE and AR data resulted in a performance decrease (71.4%) compared to the LE-only system.
• Statistical analysis indicated that mean, variance, and channel normalization are potentially beneficial preprocessing steps, while cepstral mean subtraction did not improve performance.

Conclusions:

• The choice of EEG data montage (LE vs. AR) has a significant impact on machine learning classification performance.
• Combining data from different montages without proper normalization can compromise model accuracy.
• Further investigation into statistical normalization techniques is warranted for robust EEG-based machine learning applications.